DEV Community

Cover image for Automate Candidate Routing in Your ATS: The Agent We Built
Dan
Dan

Posted on Originally published at nodesparks.com

Automate Candidate Routing in Your ATS: The Agent We Built

#automatecandidateroutingats #candidateroutingautomation #recruitingagencyautomation #atsworkflowautomation

A recruiting agency does not drown in candidates. It drowns in sorting them. Every inbound application is a small decision: which role, which recruiter, which stage, is this a knockout or a maybe. Done by hand, across several open roles and several recruiters, that decision runs hundreds of times a week and it is exactly the kind of repetitive, rules-based work a script does better. In 2026, Ashby's analysis of more than 100 million applications found applications per hire have tripled since 2021 to more than 300 on average (Ashby, 2026). The sorting problem got three times heavier while headcount did not.

This is a spoke in Cluster 2, automating your daily ops. It links up to our Ops Automation Playbook, the pillar that ranks the workflows worth attacking first. Here we go narrow and concrete: the candidate-routing agent we build for recruiting agencies. The architecture, where it works, where it misroutes and how we catch that, and the honest build-versus-buy math, including one myth you should stop repeating before you automate anything.

By Dan Colta. We are a founder-led EU automation studio. We build owned automations for SME teams, and this teardown is written from the routing builds we actually ship, not a product brochure.

Key Takeaways

  • Applications per hire have tripled since 2021 to more than 300, and candidates are roughly 50% less likely to get an interview than five years ago (Ashby, 2026).
  • The "ATS auto-rejects 75% of resumes" claim is a debunked myth from a vendor that closed in 2013 (The Interview Guys). Routing moves people forward; it should not silently bin them.
  • Route on structured, job-relevant signals and keep a human on judgment. Manual screening carries measurable bias: identical resumes with white-sounding names got ~50% more callbacks (NBER, 2004).
  • Build routing you own. 76% of recruiters expect to replace their primary recruiting platform within two years (Employ, 2025); vendor-neutral logic survives the switch.
  • Run cost is about $5/month; the real cost is the one-time build. It typically pays back inside a year against per-recruiter SaaS seats.

Why automate candidate routing now?

Because the volume has outrun the humans, and speed decides who you place. Ashby's 2026 data shows applications per hire tripled since 2021 while candidates became roughly 50% less likely to receive an interview than five years ago (Ashby, 2026). More applicants, thinner odds per applicant, same recruiter hours. Manual sorting becomes the bottleneck that makes good candidates wait.

The chart below shows the squeeze: the work per hire climbed while the interview rate fell.




There is a market signal too. The applicant tracking system market is growing from $3.28 billion in 2025 to a projected $4.88 billion by 2030, a compound annual growth rate of 8.2% (MarketsandMarkets, 2025). Recruiters are already reaching for automation: 65% say they use AI within their workflows, and 52% name getting more candidates per role as their top priority (Employ, 2025). Routing is where more candidates stops being a burden and starts being an advantage.

First, kill the 75% myth

Before automating anything, drop the idea that an ATS already auto-rejects three quarters of resumes. The claim that "75% of resumes are rejected by an ATS before a human sees them" is a marketing myth. It traces to a vendor that shut down in 2013 and has no dataset or methodology behind it, and multiple 2025-2026 analyses debunk it (The Interview Guys; HiringThing). Most ATS platforms do not enable content-based auto-rejection by default. Recruiters set knockout questions, work authorization, location, licensure, not silent resume shredding.

Why does this matter for a build? Because it sets the goal. Routing automation exists to move qualified people forward faster and to surface uncertain cases for a human, not to reject people a machine guessed were unqualified. If you build it as a rejection machine, you will bin good candidates and never know. Build it as a sorter, and you keep humans on the judgment while the software handles the busywork.

How the candidate-routing agent works

The agent is a small service that wakes on each new application, scores it against the role, and assigns it. Nothing exotic: read the applicant, apply your rules, write the result back to the ATS, notify the owner. Because the logic lives in our code rather than inside one vendor, the same agent routes across whatever ATS an agency runs.

Here is the flow.








Four stages do the work. Ingest: the agent reads each new applicant through the ATS API or an inbound webhook, so it works on Greenhouse, Lever, Ashby, Workable, or anything with an API. Knockout gate: hard, job-relevant requirements only, work authorization, location, licensure. Fail one and the candidate is held for a human with a reason logged, never silently dropped. Score against role: structured signals mapped to the role's explicit requirements, years in a named skill, specific certifications, not proxies like school prestige. Assign and notify: the agent sets the recruiter, stage, and tags in the ATS and pings the owner in Slack, the same ops-in-Slack pattern we use for reporting.

How do you route without adding bias?

Route on structured, job-relevant signals, log every decision, and send borderline cases to a human. This is not optional polish; it is the core design constraint. Manual screening already carries measurable bias. The landmark Bertrand and Mullainathan field experiment found identical resumes with white-sounding names received about 50% more callbacks than those with Black-sounding names (NBER, 2004). Any system that learns from past human decisions can absorb that pattern and repeat it at scale.

So the agent never scores on name, photo, age proxies, school prestige, or employment gaps. It scores on requirements the role actually lists. Every routing decision writes an audit record, who was routed where, on which rule, at what score, so an agency can review and correct the logic. And anything the rules cannot confidently place goes to a recruiter, not to an auto-decline. Automation removes the busywork; it must not launder discrimination or hide it behind a score.

Where it misroutes, and how we catch it

Automated routing is not magic, and pretending otherwise is how agencies get burned. Every routing build has failure modes, and naming them is part of the job. The three we see most:

Failure mode What happens How we catch it
Ambiguous role fit Candidate matches two open roles, gets sent to one, misses the better one Multi-match rule flags to human instead of picking; weekly audit of held cases
Stale role rules Requirements changed but the encoded rules did not, so good people get held Rules reviewed with the recruiter each time a role opens; version-logged
Parsing gaps A non-standard resume or CV format scores low on a real match Confidence threshold routes low-confidence parses to human review, not decline

The pattern across all three: when the agent is unsure, it escalates to a person rather than guessing. That single rule, low confidence goes to a human, is what keeps routing automation a help instead of a liability. We tune the thresholds during the first weeks against real applicants, then the hold queue shrinks as the rules sharpen.

What does it cost to build versus buy?

The running cost is trivial; the decision is one-time build versus forever-rising seats. A single routing agent runs on a roughly $5 a month server because it only fires when applications land. The real spend is the one-time build: mapping stages, encoding role rules, wiring the ATS API, and tuning thresholds. Weigh that once against per-recruiter SaaS automation seats that recur monthly and climb, the ATS market's 8.2% annual growth is priced into those seats over time (MarketsandMarkets, 2025).

Two numbers frame the payback. First, average cost-per-hire sits around $4,100 on the SHRM benchmark, so even a modest gain in speed-to-first-response protects revenue on roles you would otherwise lose to a faster agency (SHRM, via Staffing Industry Analysts). Second, and this is the vendor-neutral case, 76% of recruiters expect to replace their primary recruiting platform within two years (Employ, 2025). If your routing logic lives inside that platform, a switch throws it away. If you own it, you re-point the connectors and keep the brain. That is the same build-vs-buy logic we walk through in workflow automation examples and the data entry automation teardown: buy for simple and temporary, own for permanent and high-volume. Candidate routing, for a working agency, is permanent and high-volume.

FAQ

The five most common questions we get about automating candidate routing are answered in the schema above: what routing automation is, whether an ATS really auto-rejects 75% of resumes, how to route without adding bias, what it costs, and which platforms it works in.

The bottom line

Candidate routing is the highest-frequency, most rules-heavy decision in an agency's day, which makes it the clearest automation win once you build it honestly. Applications per hire have tripled, the odds per candidate have thinned, and the agency that responds first tends to place. Automate the sorting so recruiters spend their hours on judgment and relationships, not triage.

Build it right and the constraints are the point: route on job-relevant signals, keep a human on every uncertain case, log every decision, and own the logic so a platform switch cannot erase it. Skip the myths, name the failure modes, and the agent becomes a durable asset rather than a black box.

If you want a second read on whether your applicant volume justifies a routing build, or a scoped estimate for owning one instead of renting per-seat automation, our team can help. Start with the Ops Automation Playbook if you want the full ranking of what to automate first.

Sources (retrieved 2026-08-10):

Originally published on nodesparks.com.

Top comments (0)